Drop NumPy build dependency (#751)

Partially addresses issue: rapidsai/build-planning#82

Even though cuCIM currently `#include`s `<pybind11/numpy.h>`, the actual C++ code appears not to use NumPy. So this attempts to drop the header and the NumPy build dependency.

Authors:
  - https://github.com/jakirkham

Approvers:
  - Mike Sarahan (https://github.com/msarahan)
  - Gigon Bae (https://github.com/gigony)

URL: #751

Author: jakirkham

conda/recipes/cucim/meta.yaml

@@ -63,7 +63,6 @@ requirements:
     {% endif %}
     - cupy >=12.0.0
     - libcucim ={{ version }}
-    - numpy 1.23
     - python
     - rapids-build-backend >=0.3.0,<0.4.0.dev0
     - scikit-image >=0.19.0,<0.23.0a0
@@ -74,7 +73,7 @@ requirements:
     {% if cuda_major != "11" %}
     - cuda-cudart
     {% endif %}
-    - {{ pin_compatible('numpy') }}
+    - numpy >=1.23,<2.0a0
     - click
     - cupy >=12.0.0
     - lazy_loader >=0.1

python/pybind11/cucim_py.cpp

@@ -19,9 +19,11 @@
 
 #include <fmt/format.h>
 #include <fmt/ranges.h>
-#include <pybind11/numpy.h>
+#include <Python.h>
+#include <pybind11/buffer_info.h>
 #include <pybind11/operators.h>
 #include <pybind11/pybind11.h>
+#include <pybind11/pytypes.h>
 #include <pybind11/stl.h>
 
 #include <cucim/cuimage.h>
@@ -445,11 +447,30 @@ py::object py_read_region(const CuImage& cuimg,
     {
         py::gil_scoped_acquire scope_guard;
 
-        auto arr = pybind11::array_t<int64_t, py::array::c_style | py::array::forcecast>::ensure(location);
-        if (arr) // fast copy
+        py::object mv_obj = py::none();
+        try
         {
-            py::buffer_info buf = arr.request();
-            int64_t* data_array = static_cast<int64_t*>(buf.ptr);
+            mv_obj = py::cast<py::object>(py::memoryview(location));
+        }
+        catch (const std::exception& e)
+        {
+        }
+
+        if (!mv_obj.is_none()) // fast copy
+        {
+            py::memoryview mv(mv_obj);
+            py::buffer_info buf(PyMemoryView_GET_BUFFER(mv.ptr()), false);
+
+            if (buf.format != py::format_descriptor<int64_t>::format())
+            {
+                throw std::invalid_argument("Expected int64 array-like");
+            }
+            if (PyBuffer_IsContiguous(buf.view(), 'C') == 0)
+            {
+                throw std::invalid_argument("Expected C-contiguous array-like");
+            }
+
+            const int64_t* data_array = static_cast<const int64_t*>(buf.ptr);
             ssize_t data_size = buf.size;
             locations.reserve(data_size);
             locations.insert(locations.end(), &data_array[0], &data_array[data_size]);